Question 1

Will an agent actually use this, or will it ignore the tool?

Accepted Answer

By the author's own eval, it ignored it. The payoff case gives the agent a real task, a branch rejected as behind main, and asks for the git commands. Nothing in the prompt mentions force-pushing, but a rule recorded in an earlier session says never to force-push, so the only route to the right answer is recalling it. With the tools available and the skill installed the agent called recall zero times out of four runs. Not blocked, not erroring, just never reached for. That is why the SessionStart hook exists: it injects your rules unconditionally at startup and again after compaction.

Question 2

Can the model rewrite my notes, or quietly lose some?

Accepted Answer

No on both, and the first is structural rather than a policy. Jev returns a probability and writes no text at all, so it cannot write into your file even in principle. Sorting happens at read time rather than write time, nothing is deleted, and when the scorer cannot run — no key, no task, rate limited — you get the whole list plus a line saying why it could not sort. Recording sends nothing over the network either; only recall with a task talks out, and constraints and corrections are never part of that request.

Question 3

How accurate is the scoring, and what do the thresholds do?

Accepted Answer

Unproven, and the README says so rather than implying otherwise. A worked example of nine entries against three tasks at 1.4 seconds each shows a clearly relevant entry scoring around 0.8 with nothing else close, and one fact written two ways scoring 0.17 and then 0.70 depending on whether it said what it was or only why. The bands are exported constants, not hidden: above 0.60 you get the entry in full, between 0.30 and 0.60 one line, below that it is omitted.

Question 4

What does Jev actually get asked, and where does my data go?

Accepted Answer

One yes-or-no question per entry that is not a constraint or correction: is this still live for the task starting now, and would not knowing it cause wrong or repeated work. The surrounding design choices matter as much as the model: decisions, measurements and threads must point at something real, so anyone can check a claim instead of trusting it. Scoring goes through the Vercel AI Gateway, which the README names as the route with published zero data retention terms. To use a different scorer, write an Asker with one ask(state, questions) method and pass it to recall.

Question 5

What kinds of facts can I save?

Accepted Answer

Five kinds: constraint and correction (a rule or a fix, needing no reference), and decision, measurement, and thread, each of which must point at something real such as a commit, file, PR, or link.

Question 6

How do I install it for Claude Code?

Accepted Answer

Run claude mcp add carryforward -e AI_GATEWAY_API_KEY=$AI_GATEWAY_API_KEY -- npx -y carryforward; other MCP clients use a JSON server entry, and desktop clients should set CARRYFORWARD_PROJECT per project so they do not share one file.

Question 7

Where are my entries stored?

Accepted Answer

One JSON line per entry in ~/.carryforward/<project>.jsonl, with CARRYFORWARD_DIR to change the folder and CARRYFORWARD_PROJECT to change the file name.

Jev Carryforward

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community